Follow the result back to its evidence.

Paper and reproducibility downloads

Choose the part of the study you need. These packages are assembled from saved files, without rerunning notebooks or models. The large local handoff also contains model checkpoints and full audio; this website distributes lighter evidence packages and three attributed audio examples.

Executed notebooks

HTML views display saved outputs. Source notebooks are copied byte-for-byte. Historical notebooks remain part of the record, but do not override the corrected results. The notebook with a saved error is explicitly labeled below.

Machine-readable results

The main CSV summary retains a version field because it includes both first-correction and extended runs. Do not pool different training regimes. The website’s main comparison follows the final manuscript’s selected rows. The JSON export identifies each row’s source and uses the numerical-audit ledger.

• summary.csv
• featurefusionsummary.csv
• branchprobesummary.csv
• yearholdoutmetrics.csv
• withinunitid.csv
• withinunitsummary.csv
• globalidstrata.csv
• secondarysummaryincluding_unseen.csv
• paired_contrasts.csv

• dswp_labels.csv

• Website figure/result data

• Notebook hashes and execution inventory
• Download provenance manifest

Research reports

These pages retain the full report text and tables, including original-stage material that is superseded. Each page identifies its status. The corrected narrative in the main chapters takes precedence when interpretations differ.

Extended corrected study

• branch probes
• feature fusion
• findings
• live progress
• pretraining provenance
• priorities
• protocol
• results
• reviewer response
• source comparison
• submission notes
• supervision matrix

• year holdout

  First corrected batch

• findings

• reviewer revision audit
• revision manuscript changes
• revision protocol
• revision response map

• revision results

  Original exploration and early review

• baseline approach

• baseline framing plan
• baseline results new
• data report
• dcce literature foundations
• eda phase0
• fair comparison
• phase1 baselines
• phase2 wham probing
• phase3 dcce
• phase4 synthetic aug
• phase5 strong baselines
• professor feedback analysis
• reviewer revision audit
• revision manuscript changes
• revision protocol
• revision response map
• revision results
• sections 5 9 method experiments discussion

Reproducing the work

Read results without executing code by opening the notebook HTML views. Recalculating a metric from saved predictions is different from regenerating embeddings, refitting a probe, or training a model. The code package preserves exact experimental scripts; the evidence package supplies lightweight records rather than all runtime inputs.

Full end-to-end reproduction also requires the corresponding audio, feature caches, model checkpoints, and compatible dependencies from the complete author handoff or upstream sources. In particular, the full numerical-audit script checks checkpoint presence and training provenance: it expects the complete research directory, not only the lightweight website downloads. Environment inventories describe recorded environments, not a tested universal lockfile. Some historical paths and notebook kernels refer to the original machine.

On a complete local copy of the research project, this command recalculates the audit without training:

python3 paper_conference/qa/audit_saved_numbers.py

The manuscript source archive can be compiled using pdfLaTeX and BibTeX with its included build script. For a new training run, consult the staged READMEs and protocol reports and write new outputs separately so the original record remains intact.

Sources, attribution, and access

Audio comes from the DSWP release, distributed under CC BY 4.0; cite its associated WhAM paper. This site includes three converted mono-PCM examples, with their identifiers and metadata. The full raw-audio collection and large model checkpoints are not downloaded automatically.

The original presentation informs the visual design. Its historical scientific claims are not imported into the current comparison. This site does not publish the original conference-review PDF or invent an external repository or DOI for unpublished artifacts.

See the bibliography, numerical audit, and research record for attribution, verification scope, and the evolution of the work.